Scholarlink

<p>Inherent speckle noise in SAR images blurs image textures and affects edge detection accuracy. Traditional denoising methods can destroy critical edge information, thereby compromising the accuracy of coastline location. This paper proposes a deep segmentation framework that integrates noise-adaptive feature extraction and multi-scale boundary enhancement to achieve both denoising and detail preservation. First, learnable noise-aware convolutions dynamically suppress speckle interference. Second, a multi-scale feature branch is applied in the encoding phase to fuse global and local information, combined with spatial pyramid pooling to obtain rich context. Gradient-guided boundary attention is then used in the decoding phase to enhance details of the land-sea boundary. Finally, deep supervision is used to collaboratively optimize cross-entropy, Dice, and boundary losses, achieving unified denoising and high-precision segmentation. This study focuses on spaceborne synthetic aperture radar (SAR) images with resolutions ranging from 1 to 10&#xa0;m, covering the typical imaging capabilities of current mainstream high-resolution SAR satellites. Experimental results show that the proposed method achieves an average accuracy of over 90% in coastline identification, with a boundary error of 0.98–1.28 pixels. Even for complex coastlines, the proposed method maintains an accuracy of over 85% and an average Intersection over Union of 0.91, validating its ability to balance denoising and boundary detail preservation in the presence of strong noise.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Coastline recognition method of spaceborne SAR images based on deep learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月27日

查看全文
  • Huafeng Pan,
  • Vladimir Y. Mariano

摘要

Inherent speckle noise in SAR images blurs image textures and affects edge detection accuracy. Traditional denoising methods can destroy critical edge information, thereby compromising the accuracy of coastline location. This paper proposes a deep segmentation framework that integrates noise-adaptive feature extraction and multi-scale boundary enhancement to achieve both denoising and detail preservation. First, learnable noise-aware convolutions dynamically suppress speckle interference. Second, a multi-scale feature branch is applied in the encoding phase to fuse global and local information, combined with spatial pyramid pooling to obtain rich context. Gradient-guided boundary attention is then used in the decoding phase to enhance details of the land-sea boundary. Finally, deep supervision is used to collaboratively optimize cross-entropy, Dice, and boundary losses, achieving unified denoising and high-precision segmentation. This study focuses on spaceborne synthetic aperture radar (SAR) images with resolutions ranging from 1 to 10 m, covering the typical imaging capabilities of current mainstream high-resolution SAR satellites. Experimental results show that the proposed method achieves an average accuracy of over 90% in coastline identification, with a boundary error of 0.98–1.28 pixels. Even for complex coastlines, the proposed method maintains an accuracy of over 85% and an average Intersection over Union of 0.91, validating its ability to balance denoising and boundary detail preservation in the presence of strong noise.